WiSentry

WiFi CSI human presence & pose detection. ESP32 microcontrollers sense people by measuring how their bodies disturb WiFi radio waves (Channel State Information). A laptop runs the ML inference and a live dashboard. No cameras, no radar, no cloud - the runtime is local.

Live dashboard in simulation mode: a person walking is detected at 90% confidence — waveform disturbance, stick figure, room-map position, and event history all update at 5 Hz. More: empty room · sitting · lying

Detection tiers

Tier	Capability	Status
1	Presence (occupied / empty)	✅ working (synthetic-trained; calibrate on real data)
2	Pose (standing / sitting / lying / walking)	✅ working (synthetic-trained; calibrate on real data)
3	17-point skeleton	🧪 experimental, demo quality

Quick start - simulation, no hardware

git clone https://github.com/Arjunsk1291/wisentry.git
cd wisentry
pip install -r requirements.txt
python setup_check.py          # all lines must say [ OK ]
python main.py --simulate      # then open http://localhost:8050

You get the full live dashboard driven by a physics-based CSI simulator: a scripted person walks in, stands, sits, lies down, and leaves every 30 s. This validates the software path only. It is not evidence of real-world sensing accuracy.

Real-hardware status

The firmware and setup path are included, but this repository does not yet publish a reproduced real-room calibration result. Treat the hardware path and all real-world accuracy as work to validate, not as a completed claim.

With real hardware

2–4 × ESP32-WROOM-32 boards (~$5 each) + USB power. That's the entire BOM.

1. Flash firmware/csi_transmitter/ to one board, firmware/csi_receiver/ to the rest — step-by-step: docs/windows_setup.md / docs/ubuntu_setup.md
2. Place them per docs/placement_guide.md
3. python main.py

How it works

ESP32 TX ──100 pkt/s──> air (person disturbs multipath) ──> ESP32 RX(s)
ESP32 RX ──UDP wire protocol v1──> laptop
laptop:  udp_server → csi_parser → signal_processor (Hampel, Butterworth,
         band features) → CNN models / rule fallback → debounced detector
         → Dash dashboard (7 panels, 5 Hz)

• Wire protocol v1 is pinned byte-for-byte across firmware, simulator, and parser (engineering specification) with a shared test vector.
• Training = runtime: models/train_all.py generates data by pushing simulator physics through the same SignalProcessor used live.
• Honest metrics: shipped weights are synthetic-trained (saved/metrics.json is tagged "data": "synthetic"); collect your own data with python main.py --collect --label standing to calibrate.

Documentation

Doc	What's in it
docs/user_manual.md	10 chapters, unboxing → live dashboard
docs/hardware_bom.md	exact parts, prices, where to buy
docs/placement_guide.md	room diagrams, coverage tables
docs/windows_setup.md	Windows + Arduino IDE flashing, every click
docs/ubuntu_setup.md	Ubuntu differences + arduino-cli path
docs/troubleshooting.md	30 symptoms with fixes
ENGINEERING_SPEC.md	full engineering specification
PROJECT_LOG.md	dated ledger of every test, success, and failure

Development

python -m pytest tests/        # unit tests (parser, DSP, detector, simulator)
python models/train_all.py     # retrain all three models
python tests/gate_phase3.py --spawn   # end-to-end dashboard gate

Project history, including what failed and why, lives in PROJECT_LOG.md. CI runs the test suite on Python 3.10 and 3.11. Model-dependent tests skip when synthetic-trained weights are absent; the status is reported rather than treated as a verified model result.